How the numbers work

We suggest the kinds of questions a buyer asks before choosing: recommendations, comparisons, alternatives, use cases, budget, and trust. You can edit them or add your own.

We query ChatGPT, Claude, Gemini, and Perplexity, plus Grok on Grow and Agency plans. We keep two kinds of answer separate:

• From memory: what the model says without searching the live web.
• After a web search: what it says with web search turned on. We run this on ChatGPT, Claude, Gemini, and Perplexity.

We only check Grok from memory. Its live search can run a long, expensive chain of searches for one answer, so we leave it out.

These answers often disagree, so we label every answer with how it was produced. Every paid plan includes the four core tools. There are no per-tool add-ons for them.

Paid plans repeat the same questions and save every answer. That makes a later audit comparable with an earlier one. We collect one answer per question, AI tool, and mode.

Running every expensive web-search check every day would waste money without teaching you much. Here's the schedule:

• The first 25 active questions get the daily pulse and the full weekly audit.
• Claude with live search is the most expensive check we run, so each week it covers a rotating group of priority questions instead of all of them. Every priority question gets a fresh Claude live answer every few weeks, and every answer is labeled with the run it came from.
• The rest of your questions get the full multi-tool scan in the first deep run of each month.
• Live re-checks let you refresh the live-search answers for one question on demand: when you ship a fix, when an alert fires, or whenever you press the button. Plans include a monthly amount.

Every stored answer has a date and mode. A score only uses the valid answers that audit actually collected.

We only record a numeric rank when the answer contains an explicit ordered recommendation list that includes your brand. A brand mentioned in prose is a mention, not a rank. Our headline positioning metrics are share of voice and top-3 rate across the complete tracked question set, rather than pretending one ordinal position is stable.

A model extracts the same fields from every answer: brand mention, rank when there is a real ordered list, sentiment, cited sources, and factual errors. Empty, cut-off, and errored answers stay in the archive but don't count toward the metrics.

The 0-100 score combines mention rate, rank, citation rate, competitor pressure, and sentiment. The dashboard shows those parts next to the score. If we change the formula, we recalculate old audits so the trend still means something.

Every recommended fix keeps the questions it was meant to affect. When you mark the work as shipped, we save the mention rate on those questions and check them again in later full audits. We also measure your other tracked questions as a comparison set, because AI answers can move even when you change nothing.

If the targeted questions improve more than the comparison set, we show that difference. If both move together, we say the result is unconfirmed. This is useful evidence, but it is not a randomized experiment and does not prove that one change caused the result.

• AI answers change constantly. Treat week-to-week movement inside the stability band as directional, not precise. We label it that way.
• We measure a defined question set, not the infinite space of phrasings. Broader coverage means more questions, which higher plans allow.
• Personalization, region, and model version affect what any individual user sees. We query neutrally and report the aggregate.
• Visibility is a leading indicator, not revenue. We tie it to referral signals and to the actions you ship, but we will not pretend a citation is a customer.